Data driven optimization of online markets like eBay, Uber, Amazon, and Zillow requires counterfactual prediction, using data from the original market, to evaluate the performance of changes to market rules and parameters. Standard methods from econometrics allow such prediction when the market place is in equilibrium. However, as is evidenced by widely fluctuating prices, bids, and behaviors, these online market places are rarely in equilibrium. Nekipelov, Syrgkanis, Tardos, (2015) demonstrated that the equilibrium assumption can be replaced with an (approximate) “no regret” assumption that, for example, is satisfied by standard machine learning algorithms. This talk will review this methodology and showcase its application to the analysis of preferences of players, market revenues, and market welfares for two markets: the display advertising auction on Zillow.com and the Amazon marketplace.